Exploring transferable and robust adversarial perturbation generation across network hierarchy

Ruikui Wang, Yuanfang Guo, Ruijie Yang, Yunhong Wang

Abstract: Highlights•We propose a black-box framework enhancing both transferability and robustness of adversarial examples.•A dynamic mechanism optimizes directional guidance for more effective intermediate attacks.•A new SOTA black-box attack successful rate is achieved by our proposed TRAP method.